Web Agent Protocol

Overview

The Web Agent Protocol (WAP) is a standardized framework designed to enable seamless interaction between users, web agents, and browsers by recording and replaying browser actions. It separates the concerns of action recording and execution, allowing for efficient automation and reusability. The Python SDK for WAP implements the full specification, making it easy to:

1. Collect user‑interaction data with the OTA‑WAP Chrome extension.
2. Convert the raw event stream into either exact‑replay or smart‑replay action lists.
3. Convert recorded actions into MCP servers for reuse by any agent or user
4. Replay those lists using the WAP-Replay protocol to ensure accurate browser operations.

WAP FULL DEMO

Without WAP

WAP Record

WAP Replay

Example using WAP

Setup

Install the dependencies with the following command:

Create a conda env

conda create -n WAP python=3.11

Activate the conda env

conda activate WAP

Install the dependencies

pip install -r requirements.txt

Setup your repo source path:

set PYTHONPATH=C:/path/to/webagentprotocol # for Windows
export PYTHONPATH=/path/to/webagentprotocol # for Linux

Create .env file under the repo root directory with your own API keys:

OPENAI_API_KEY=sk-proj-...
DEEPSEEK_API_KEY=sk-...

Record

WAP record extension

Please refer to OTA‑WAP Chrome Extension to setup action capturer in your Chrome browser.

Start data‑collection server

Run the following command to start the server to collect data from the extension:

python action_collect_server.py

Once the server is up, you can start to record from the page using WAP Chrome extension.

The server listens on http://localhost:4934/action-data by default, please make sure the Host and Port in the extension settings match this server config. Each session will be saved to:

data/YYYYMMDD/taskid/summary_event_<timestamp>.json

An example of the formatted data which you will received in the WAP backend server is like:

{
  "taskId": "MkCAhQsHgXn7YgaK",
  "type": "click",
  "actionTimestamp": 1746325231479,
  "eventTarget": {
    "type": "click",
    "target": "<a ota-use-interactive-target=\"1\" data-ordinal=\"3\" href=\"https://www.allrecipes.com/recipe/68925/cheesy-baked-salmon/\" data-tax-levels=\"\" data-doc-id=\"6592066\" class=\"comp mntl-card-list-card--extendable mntl-universal-card mntl-document-card mntl-card card card--no-image\" id=\"mntl-card-list-card--extendable_3-0\">\n



...",
    "targetId": "mntl-card-list-card--extendable_3-0",
    "targetClass": "comp mntl-card-list-card--extendable mntl-universal-card mntl-document-card mntl-card card card--no-image"
  },
  "allEvents": {},
  "pageHTMLContent": "<header data-tracking-container=\"true\" data-collapsible=\"true\" class=\"comp header mntl-header mntl-header--magazine mntl-header--open-search-bar mntl-header--myr\" id=\"header_1-0\"><a data-tracking-container=\"true\" id=\"mntl-skip-to-content_1-0\" class=\"mntl-skip-to-content mntl-text-link\" rel=\"nocaes\" href=\"#main\"></a>

..."
}

Generate replay lists

Mode	Command
Exact replay – exactly reproduce every action	python wap_replay/generate_exact_replay_list.py --data_dir_path data/<date>/<task_id> --output_dir_path data_processed/exact_replay
Smart replay – condensed goal‑oriented steps	python wap_replay/generate_smart_replay_list.py --data_dir_path data/<date>/<task_id> --output_dir_path data_processed/smart_replay

Replace with the folder produced by the extension (e.g. em3h6UBDZykz0gnH).

Output structure:

data_processed/smart_replay/
 ├─ subgoals_<task_id>/                     # intermediate prompts & replies
 └─ wap_smart_replay_list_<task_id>.json   # final smart replay list for the agent

data_processed/exact_replay/
 └─ wap_smart_replay_list_<task_id>.json   # final exact replay list for the agent

Replay

python run_replay.py --model-provider openai --wap_replay_list data_processed/exact_replay/wap_exact_replay_list_<task_id>.json --max-concurrent 1

For smart-replay, replace the path with a smart‑replay JSON to test this mode.

Convert to MCP Server

python wap_replay\generate_mcp_server.py --task_id <task_id>

converted MCP servers will be located under mcp_servers folder

Replay with MCP

You would need 2 terminals to replay with MCP. In the first termnial

python wap_service.py

In the second termnial

python mcp_client.py

Then enter your prompt in the second terminal

example: find a top rated keyboard on amazon.ca using smart replay

Troubleshooting

ModuleNotFoundError – run commands from the project root or export PYTHONPATH=. (set PYTHONPATH=. for Windows).

“no task‑start file” – ensure the extension recorded a full session; the generators require exactly one task-start and one task-finish record.

Acknowledgement

Browser-Use: https://github.com/browser-use/browser-use

MCP: https://github.com/modelcontextprotocol/python-sdk

DOM Extension: https://github.com/kdzwinel/DOMListenerExtension